Optical Burst Switching (OBS)

Abstract
In this seminar, we give an introduction to optical burst switching (OBS) and compare
it with other existing optical switching paradigms. Basic burst assembly algorithms and
their effect on assembled burst traffic characteristics are described first. Then a brief
review of the early work on burst transmission is provided followed by the description
of a prevailing protocol for OBS networks called Just-Enough-Time (JET). Algorithms
used at an OBS core node for burst scheduling as well as contention resolution strategies
are presented next. Tradeoffs between their performance and implementation
complexities are discussed. Recent work on QoS support, IP/WDM multicast, TCP
performance in OBS networks and Labelled OBS is also described, and several open
issues are mentioned.

Contents
1 Introduction
2 OBS Fundamentals
3 Qualitative Comparison
4 Burst Assembly
o Assembly algorithms
o Assembled burst traffic characteristics
5 Burst Reservation Protocols
6 Burst Switching
o Scheduling Algorithms
o Contention Resolution
7 Towards an Optical Internet
Service differentiation
IP/WDM Multicast and TCP over OBS
LOBS
8 Concluding Remarks
9 References

1 Introduction
With recent advances in wavelength division multiplexing (WDM) technology, the amount of raw
bandwidth available in fiber links has increased by many orders of magnitude. Meanwhile, the
rapid growth of Internet traffic requires high transmission rates beyond a conventional electronic
router's capability. Harnessing the huge bandwidth in optical fiber cost-effectively is essential for
the development of the next generation optical Internet.
Several approaches have been proposed to take advantage of optical communications and in
particular optical switching. One such approach is optical circuit switching based on wavelength
(A) routing whereby a lightpath needs to be established using a dedicated wavelength on each link
from source to destination. Once the connection is set up, data remains in the optical domain
throughout the lightpath. An alternative to optical circuit switching is optical packet switching. In
optical packet switching, while the packet header is being processed either ail-optically or
electronically after an Optical/Electronic (O/E) conversion at each intermediate node, the data
payload must wait in the fiber delay lines and be forwarded later to the next node.
In order to provide optical switching for next generation Internet traffic in a flexible yet feasible
way, a new switching paradigm called optical burst switching (OBS) was proposed in. Various
OBS approaches with different tradeoffs have since been described. There are two common
characteristics among these variants:
Client data (e.g., IP packets) goes through burst assembly/disassembly (only) at the edge of
an OBS network,nevertheless, statistical multiplexing at the burst level can still be achieved
in the core of the OBS network.
Data and control signals are transmitted separately on different channels or wavelengths (A's)1,
thus, costly O/E/Oconversions are only required on a few control channels instead of a large
number of data channels.
In this tutorial, we first introduce the basic idea of OBS, compare it with other switching
paradigms and point out why OBS is a viable technology for the next generation optical Internet.
Techniques for generating a burst at the edge
Hereafter, we will use the terms channel and wavelength (or A) interchangeably of an OBS

network are studied next, followed by discussions on various burst reservation protocols.
Characteristics of the burst traffic assembled using different assembly algorithms are also
analyzed. The subsequent section focuses on issues at a core OBS node: namely burst
scheduling and contention resolution. Recent efforts on supporting service differentiation,
IP/WDM multicast, the performance of TCP and other issues related to traffic engineering in
OBS networks are described at the end.

2 OBS Fundamentals

Figure 1: Burst Assembly/Disassembly at the Edge of an OBS Network

In an OBS network, various types of client data are aggregated at the ingress (an edge
node) and transmitted as data bursts (Figure l(a)) which later will be disassembled at the egress
node (Figure l(b)). During burst assembly/disassembly, the client data is buffered at the edge where
electronic RAM is cheap and abundant.

Figure 2: Separated Transmission of Data and Control Signals

Figure 2 depicts the separation of data and control signals within the core of an OBS network.
For each data burst, a control packet containing the usual "header" information of a packet including
the burst length information is transmitted on a dedicated control channel. Since a control packet is
significantly smaller than a burst, one control channel is sufficient to carry control packets

associated with multiple (e.g., hundreds of) data channels. A control packet goes through O/E/O
conversion at each intermediate OBS node and is processed electronically to configure the
underlying switching fabric. There is an offset time between a control packet and the
corresponding data burst to compensate for the processing/configuration delay. If the offset time is
large enough, the data burst will be switched ail-optically and in a "cut-through" manner, i.e.,
without being delayed at any intermediate node (core). In this way, no optical RAM or fiber delay
lines (FDLs) are necessary at any intermediate node. Nevertheless, the burst-level granularity leads
to a statistical multiplexing gain, which is absent in optical circuit switching. Furthermore, it
allows a lower control overhead per bit than that in optical packet switching as to be discussed next
in more detail.

3 Qualitative Comparison

Figure 3: Comparison of Different Switching Node Architectures

Basic switching node architectures used by various switching paradigms are illustrated and
compared in Figure 3. At an optical circuit switching (OCS) node shown in Figure 3(a), once a
lightpath is setup, all data carried by one input A will go to a specific output A. Since no O/E/O
conversion of data at any intermediate node is needed, multi-hop transparency (in terms of the bit
rate, protocol and coding format used) can be achieved. On average, the connection duration

should be on the order of minutes or longer as setting up or releasing a connection takes at least a
few hundreds of milliseconds. Shorter duration connections needed to accommodate sporadic data
transmissions will result in a prohibitively high control overhead. A major difference between OCS
and the other three approaches depicted in Figure 3 is that in OCS, no statistical multiplexing of the
client data can be achieved at any intermediate node. More specifically, in the core, bandwidth is
allocated by one A at a time, which is a coarse granularity. In practice, however, most of today's
applications only need the sub-A connectivity. In addition, high-bit rate computer communications
often involve "bursts" that last only a few seconds or less.
To overcome the above deficiency of the OCS approach, O/E/O conversion can be introduced
above an OCS network in the IP and SONET layers for example. The electronic switching node
used in such an O/E/O approach is depicted in Figure 3(b). Here, statistical multiplexing of the
client data at the sub-A granularity is possible with electronic processing and buffering (not
shown in Figure 3(b)). Since every data unit needs to go through O/E and E/O conversion, this
approach is not scalable enough to support hundreds of wavelengths, each working at 40Gbps or
beyond (the need for which is anticipated in the near future). In addition, electronic switches are
known to suffer from problems such as limited capacity and huge power/space consumption and
heat dissipation in addition to requiring expensive O/E/O conversions. Note that, although not
shown, either an optical cross connect or optical add-drop multiplexor may also be used in
conjunction with an electronic switch for wavelength granularity traffic that does not need to go
through the electronic switch. A hybrid, multi-layer network consisting of such nodes, each
consisting of both an electronic switch/router and an optical cross connect, is one way to combine
the strength of the optics and electronics, but certainly not the only way to do so, and in fact may
not be the ultimate long-term solution.
Since all-optical header processing will not be economically viable in the near future due to the
immaturity of high-speed optical logic, the optical packet switching (OPS) approach will likely
require each header to go through O/E conversion for processing and E/O conversion for
transmission (Figure 3(c)). An important difference from the previous O/E/O approach is that here,
the header can potentially be sent at a much slower rate than the data using for instance sub-carrier
multiplexing, thereby easing the speed requirement on the O/E/O conversion devices while still
maintaining a high data throughput. Nevertheless, OPS is difficult to implement because of its need
for a large number of O/E/O conversion devices (one set for each wavelength), header
extraction/insertion mechanisms (though not shown in the figure) as well as FDLs and packet

synchronizers. Note that, an optical cross-connect or add-drop multiplexor mentioned above can
also be used in conjunction with the OPS nodes or OBS nodes to be discussed below if/when it is
more economic to do so.
In the OBS paradigm, only a few control channels (e.g., one per fiber) go through O/E/O
conversion (see Figure 3(d)). Given that the data is switched ail-optically at burst level, data
transparency and statistical multiplexing can be achieved concurrently. Since OBS takes advantage
of both the huge capacity in fibers for switching/transmission and the sophisticated processing
capability of electronics, it is able to achieve cost reduction and leverage the technological
advances in both optical and electronic worlds, which makes it a viable technology for the next
generation optical Internet.
At an OBS node, no synchronization/alignment of bursts is necessary unless the switching fabric
operates in a slotted manner. In addition, FDLs and wavelength converters which are optional can
help in reducing burst loss. Currently, it is a challenge to implement an OBS switching fabric with
hundreds of ports operating at a switching speed which is on the order of nanoseconds.
Nevertheless, on-going research work has shown promise.

4 Burst Assembly

Figure 4: Architecture of an OBS Ingress Node

Burst assembly is the procedure of aggregating packets from various sources, such as an IP
router, into bursts at the edge of an OBS network. The architecture of a typical OBS ingress node
is shown in Figure 4. The switching unit forwards incoming packets to burst assembly units. The
packets to the same OBS egress node are processed in one burst assembly unit. Usually, there is
one designated assembly queue for each traffic class (or priority). The burst scheduler is in charge
of creating bursts and their corresponding control packets, adjusting the offset time for each burst,
scheduling bursts on each output link and forwarding the bursts and their control packets to the
OBS core network .
Recent studies on burst assembly have shown that different assembly schemes affect the
assembled burst traffic's characteristics.
4.1

Assembly algorithms

Usually, assembly algorithms can be classified as timer-based, burstlength-based and mixed

timer/burst length-based ones.
In the timer-based scheme, a timer starts at the beginning of each new assembly cycle. After a
fixed time T, all the packets that arrived in this period are assembled into a burst. In the
burstlength-based scheme, there is a threshold on the (minimum) burst length. A burst is assembled

when a new packet arrives making the total length of current buffered packets exceed the threshold.
The time out value for timer-based schemes should be set carefully. If the value is too large, the
packet delay at the edge might be intolerable. If the value is too small, too many small bursts will be
generated resulting in a higher control overhead. While timer-based schemes might result in
undesirable burst lengths, burstlength-based assembly algorithms do not provide any guarantee on
the assembly delay that packets will experience. To address the deficiency associated with each
type of the assembly algorithms mentioned above, mixed timer/threshold-based assembly
algorithms were proposed in. For example, in, a burst can be sent out when either the burst length
exceeds the desirable threshold or the timer expires.
Adaptive assembly algorithms were also proposed to optimize the performance of OBS networks
in which either the time threshold or the burst length threshold or both are adjusted dynamically
according to real time traffic measurements. They provide better performance especially with
strongly correlated input packet traffic but have a higher operational complexity.
After a burst is generated using the algorithms mentioned above, the burst is buffered in the
queue for an offset time before being transmitted to give its corresponding control packet enough
time to make reservations at the downstream nodes as shown in Figure 2. During this offset period,
packets may continue to arrive. Including those packets in the same burst is usually unacceptable
because the reservation at the downstreams nodes may have already been made based on the original
burst length record in the control packet. Leaving those packets for the next burst on the other hand,
will increase the average delay especially when the traffic load is heavy. One way to minimize this
extra delay is to perform burst length prediction: let the control packet carry a burst length of I +
f(t) instead of I, where I is the exact burst length when the control packet is sent out, and f(t) is the
predicted extra burst length as a result of additional packet arrivals during the offset time t.
Assume that the total length of packets actually arriving during the offset time is l(t). If l(t) < f(t),
part of bandwidth reserved will be wasted. Otherwise (i.e., if l(t) > f(t)\ only a few extra packets
(whose total length is about l(t) /(t)) are delayed to be transmitted in the next burst.
4.2

Assembled burst traffic characteristics

Recently, the sensitivity of OBS network performance to the assembled burst traffic characteristics
such as inter-arrival time and burst length distribution has been studied, these studies have focused
on the statistical characteristics of burst traffic, which can be divided into two categories: short

range (small time scales) and long range (large to infinite time scales). In most of these studies, the
packet arrivals into an assembly queue from many independent traffic sources were assumed to be
Poisson. For a timer-based assembly algorithm, the size of a burst is equal to the sum of the size of
all the packets arriving in a fixed time period, and was shown in to be a Gaussian distributed
random variable according to the central limit theorem. Conversely, for a burstlength-based
assembly algorithm, the burst inter-arrival times have a Gaussian distribution. Similar observations
were reported via simulation in. The general conclusion is that the short range burstiness in the input
packet traffic is alleviated due to burst assembly and the smoothed assembled burst traffic can
enhance the network's performance.
An important characteristic of today's Internet traffic is its long range dependency, which
increases data loss and delay and decreases network resource utilization in electronic packet
switched networks. Although it was claimed in that burst assembly algorithms could reduce the
long range dependency in the input IP packet traffic, pointed out that long range dependency in the
traffic will not change after burst assembly. On the other hand, the results in [14] showed that the
influence of the long range dependency on the performance of an OBS node (i.e., in terms of burst
loss rate) is negligible because of its bufferless nature.
If a timer-based assembly scheme is used, the bursts' inter-arrival time will be a constant.
Furthermore, if a burstlength-based assembly algorithm is used, the variance of the inter-arrival
time of the bursts coming from different edge nodes may become small when the traffic load is
heavy. In such cases, undesirable persistent collisions of bursts from different sources might happen
if these sources are adversely synchronized. Adding a randomized extra offset time at each edge
node may prevent such synchronization among the sources.

5 Burst Reservation Protocols

Although the concept of burst switching was introduced for centralized TDMA systems and ATM
networks in early 1990, protocols suitable for high speed WDM optical networks were not
developed until 1997. In this section, we will first give an overview of early burst transmission
protocols followed by an introduction to protocols for OBS networks.
It is evaluated two burst level admission control mechanisms for ATM networks: tell-and-wait
and tell-and-go. In the former, when a source has a burst to transmit, it first tries to reserve the
bandwidth/wavelength from the source to its destination by sending a short 'request' message.
Every intermediate node receiving this message will make a reservation on a specific output link.
If the requested bandwidth is successfully reserved on all the links along the path, an ACK will be
sent back to inform the source to send out the burst immediately; Otherwise, a NAK will be
returned to release the previously reserved bandwidth, and initiate the retransmission of the
'request' message after a backoff time. In tell-and-go, on the other hand, the source transmits bursts
without making any bandwidth reservation in advance. At an intermediate node, the burst needs to
be delayed before the switch control unit makes an appropriate reservation on an outgoing link. If
the reservation fails at any intermediate node, a NAK will be sent back to the source to initiate the
retransmission of the burst after a backoff time.
Various performance comparisons (in terms of e.g., throughput and delay) between these two
conceptual approaches were given in. It has been found that tell-and-wait outperforms tell-and-go
when the propagation delay is negligible with respect to the burst length. The opposite becomes true
when the propagation delay is significant compared with the burst length.
The concept of tell-and-go forms the basis of Terabit Burst Switching. With this approach, in order
to compensate for the control packet processing time and prevent a burst from entering the
switching fabric before its configuration is finished, a fixed delay is inserted into the data path using
a FDL at each input port. On the other hand, Just-In-Time (JIT), which was first proposed in, can be
considered as a variant of tell-and-wait as it requires each burst transmission request to be sent to a
central scheduler. The scheduler then informs each requesting node the exact time to transmit the
data burst. Here, the term Just-In-Time means that by the time a burst arrives an intermediate node,
the switching fabric has already been configured. This concept was later applied and extended to a
Wavelength Routed OBS network. Since centralized protocols are neither scalable nor robust,

provided a distributed version of JIT protocol called Reservation with Just In Time, which requires
a copy of the request to be sent to all switches (each has a scheduler) concurrently. These schedulers
are not only synchronized in time, but also share the same global link status information, which
makes the implementation difficult. The authors of proposed another distributed version of the JIT
protocol based on hop-by-hop reservation which adopts some features of the Just Enough Time
(JET) protocol.
JET is the most prevailing distributed protocol for OBS networks today which does not require
any kind of optical buffering or delay at each intermediate node. It accomplishes this by letting
each control packet carry the offset time information and make the so called delayed reservation for
the corresponding burst, i.e., the reservation starts at the expected arrival time of the burst. In the
example shown in Figure 5, the bandwidth is reserved for the first burst starting

Figure 5: JET Protocol

from the burst arrival time instead of the arrival time of control packet. At each intermediate node,
the offset time is updated (reduced) to compensate for the actual control packet processing/switch
configuration time (see Figure 2). Note that the delay experienced by a control packet might vary for
different reasons. In addition, when its considered deflection routing in an OBS network, the
minimal offset time for the primary path might not be enough if the burst takes a longer alternate
path. In such a case, an extra offset time can be added .
Another important feature of JET is that the burst length information is also carried by the
control packet, which enables it to make a closed-ended reservation (i.e., only for the burst duration
with automatic release) instead of an open-ended reservation (i.e., which would not be terminated
until a release signal is detected). This closed-ended reservation helps the intermediate node make

intelligent decisions as to whether it is possible to make a reservation for a new burst and thus the
effective bandwidth utilization can be increased. An example is shown in Figure 5 where the
reservation for the 2nd burst arrival in Cases 1 and 2 can succeed if and only if at the time when the
2nd control packet arrives, the intermediate node makes closed-ended reservations for both the first
and second bursts.

6 Burst Switching
In a conventional electronic router/switch, contention between packets can be resolved by
buffering. However, in OBS networks, no or limited buffering is available and thus burst
scheduling and contention resolution must be done in a different manner.
6.1 Scheduling Algorithms
When wavelength conversion capability is assumed, an incoming burst may be scheduled onto
multiple wavelengths at the desired output port. A burst scheduler will choose a proper wavelength
for this burst taking into consideration the existing reservations made on each wavelength, and
make a new reservation on this selected channel. Below, we will describe several scheduling
algorithms.
The scheduling horizon is defined as the latest time at which the wavelength is currently
scheduled to be in use. In Figure 6, for example, time t'{ is the scheduling horizon for channel
Ci. A simple scheduling algorithm: Horizon [5], which is also called the LAUC (latest available
unscheduled channel) algorithm in [7] works as follows, for each wavelength, a single scheduling
horizon is maintained. Only the channels whose scheduling horizons precede the new burst's arrival
time are considered "available" and the one with the latest scheduling horizon is chosen. The
horizon is then updated after making the reservation for the next burst. The basic idea for this
algorithm is to minimize bandwidth gaps/voids created as a result of making a new reservation. In
Figure 6, channel C3 will be reserved if Horizon is applied.
Simplicity in both operation and implementation is the main advantage of the Horizon-based
algorithms. However, they waste the gaps/voids between two existing reservations, e.g., t( ti on
channel Ci in Figure 6. When a FDL set is available or the offset-time based QoS [19] scheme to be
mentioned in the following section is applied, many such voids will be generated. Therefore,
algorithms capable of void filling, i.e., making new reservations within existing gaps are

desirable. For example, using LAUC-VF (LAUC with void filling) proposed in, channel C\ will be
chosen.
Several variants of the LAUC-VF algorithm including Min-SV (Starting Void), Min-EV(Ending
Void) and Best Fit were proposed in [20]. Min-SV is functionally the same as LAUC-VF but a
much faster implementation is achieved using a technique from computational geometry. On the
other hand, Min-EV tries to minimize the new void generated between the end of new reservation
and an existing reservation while Best Fit tries to minimize the total length of starting and ending
voids generated after the reservation. Figure 6 illustrates the outcomes of these three scheduling
algorithms.
I Algorithms I Time complexity I State information I Bandwidth Utilization I

Table 1:

Comparison of Different scheduling Algorithms

The performance of various scheduling algorithms was compared in which shows that LAUCVF, Min-SV, Min-EV and Best Fit have a comparable bandwidth utilization (or loss rate) which is
much higher (or lower) than Horizon based algorithms. The running time complexity of different
scheduling algorithms was also analyzed. Table 1 summarizes the above discussion using the
following notations:

W: Number of wavelengths at each output port

M: Maximum number of data bursts (or reservations) on all channels

Horizon-i'. Horizon of the ith data channel

Sitj and Eitj-. Starting and ending time of jth reservation on channel i

From Table 1, the Min-SV/EV algorithms are the most desirable among all void-filling algorithms.
In fact, one can minimize the void newly generated by first searching for a proper void using MinEV first, and then if (and only if) such a proper void cannot be found, search for a horizon using
Min-SV.

6.2 Contention Resolution

Using one way reservation protocols such as JET, the ingress node sends out bursts without having
reservation acknowledgements or global coordination. This however, requires an intermediate
OBS node to resolve possible contention among bursts. In a bufferless OBS network, contention
among the bursts can be resolved in three ways: deflection, dropping and preemption.
Through deflection, a burst is sent to a different output channel instead of the preferred one.
Since contention can only happen when bursts compete for the same wavelength on the same

output port simultaneously, deflection can be applied in wavelength, space and/or time domains.
Wavelength domain: a contending burst can be sent on another wavelength through
wavelength conversion.
Space domain: a contending burst can be sent to a different output port and then follow an
alternate route to the
destination [4].
Time domain: by passing through an FDL, a contending burst can be delayed for a fixed
time.
If a contending burst cannot be deflected due to the unavailability of any wavelength, output
port or FDL, data loss becomes inevitable. More specifically, a common approach is to drop the
incoming burst (which is a non-preemptive approach). In addition, it is possible for the incoming
burst to preempt an existing burst based on priority or traffic profile. It is also possible to break
the incoming burst or the existing burst into multiple segments, and each segment can then be
deflected, dropped or preempted. This approach was called burst segmentation in and OCBS in.

Table 2: Comparison of Different Contention Resolution Schemes

Table 2 gives a brief summary of these contention resolution schemes. Note that some of these
contention resolution schemes can be applied jointly. For example, instead of simply forwarding a
burst onto an alternate route (using deflection routing) when contention happens, one can deflect a
burst along a pre-determined path that returns the burst to the node where the deflection occurred
and then forwards it along the original route. With this approach, the network acts like a buffer (or
FDL).
Unlike all the contention resolution schemes mentioned above which work in a passive

manner, i.e., taking certain actions after a contention occurs, one may collect the burst loss
performance statistics on different wavelengths and rank them with priorities accordingly. Bursts
are then assigned to higher priority wavelengths which have lower burst loss rates whenever
possible [24]. However, this approach can only be taken by edge nodes in a network without
wavelength converters. We observe that one can also pro-actively reduce burst contention (and
loss) by using either the electronic buffer at an edge node or FDLs at an upstream node to
sequentialize the bursts on as few wavelengths as possible so as to reduce the number of bursts
which might cause overlapped reservations on an output channel at a downstream node.

7 Towards an Optical Internet

Over the past few years, running IP applications directly above the optical layer has received a
considerable amount of attention. In this section, we give an overview of several initial steps
toward building an optical Internet based on the OBS paradigm.

7.1 Service differentiation

Much research work has been devoted to QoS provisioning in the Internet. However, various
Internet service disciplines and packet scheduling algorithms developed in the literature are based
on electronic packet switching and mandate the use of buffers. In the optical domain, a FDL can
provide a limited and deterministic delay but it is incapable of providing most of the buffer
management functions as an electronic RAM does. To address the discrepancy between a bufferless
OBS network and an electronic packet switched network, three different approaches have been
proposed to provide service differentiation in OBS networks.

Arrival time of new control packet

-Offset 1>-class 1 (low priority)
Offset 2------>-class 2 (high priority)
Figure 7: Offset Time's Effect on Burst Loss
The first approach manages QoS on a class by class basis using different extra offset times for
different classes of bursts . The basic idea is by giving a larger extra offset time to a higher
priority class, reservation for a higher priority burst can be made in much advance than lower
priority bursts and thus has a better chance to succeed. Figure 7 shows that a long offset time enables

a high priority class burst to succeed in making a reservation. Studies have shown that the
probability that a low priority burst will block a high priority burst can be negligible when the
difference of offset time between these classes is a few times of the average burst length of the low
priority class .
Although the offset time based differentiation is easy to implement and provides efficient
isolation between service classes when a sufficiently large extra offset time is assigned to higher
priority bursts, the extra offset time introduces an additional delay at the edge and in addition, the
performance of the differentiation depends on the burst length and inter-arrival time distributions.
Active dropping was thus proposed to avoid the shortcomings mentioned above. In this alternative
to the offset time based differentiation, selective dropping of bursts is initiated according to either
loss rate measurement or traffic profile to guarantee that the high priority class will have a better
chance to make successful reservations.
While the above two approaches can provide differentiation at the burst level, differentiation at
the packet level can be achieved with burst segmentation. In such an approach, packets from
different service classes are assembled into different bursts. When contention occurs, low priority
bursts will be segmented and experience a higher packet loss probability. Alternatively, instead of
assembling a burst with packets from a single service class only, packets from low priority service
classes can be assembled to form the tail or head of each burst, whereas packets from high priority
service classes are assembled in the middle of each burst. If segments at the tail or head of a burst
are dropped when contentions happen, differentiation on packet loss can be achieved.
7.2 IP/WDM Multicast and TCP over OBS
Many of today's and emerging Internet applications can be more efficiently supported using
multicast. A straightforward way to do multicasting in an OBS network is Separate Multicast in
which multicast traffic and unicast traffic are assembled separately into different bursts. To reduce
the overhead due to guard bands and control packet associated with each burst, a scheme called TreeShared Multicasting was proposed, whereby multicast traffic belonging to different multicast
sessions can be assembled together in a burst, which is then delivered via a shared multicast tree.
Various criteria for determining whether two multicast sessions should share a tree and various
algorithms for constructing the shared multicast trees were presented in. Since it is possible that
some data in a burst is delivered to non-intended destinations via a shared multicast tree, the benefit
of multicast sharing strategy depends on the degree of overlapping among the multicast sessions

that share the multicast tree.

OBS networks also have unique characteristics that affect TCP throughput performance.
Since TCP is the most widely used protocol for data transmissions, understanding TCP
performance in an OBS network, which may become a future Internet backbone and support a large
amount of TCP traffic, is thus of much interest. Several recent studies have investigated the
interactions between OBS and TCP congestion control mechanisms. For example, the study in
found that the burst assembly process introduces a delay penalty in TCP throughput because it
increases the round trip time, which in turn decreases TCP throughput. On the other hand, the
enlarged transmission unit from a packet to a burst can increase the amount of data sent between
two losses, resulting in the so-called "correlation gain". More specifically, a TCP source with a
relative low access bandwidth in the local IP access network and a small burst assembly time at the
edge can have only one TCP segment in each burst, and thus there is no correlation gain. But since
the delay penalty is also insignificant in this case, the throughput is similar to that without burst
assembly. For a TCP source with a relatively high access bandwidth and a large burst assembly
time, all TCP segments from one sending window can be assembled into one burst, and hence, the
correlation gain is maximized but the delay penalty is also large. In our recent studies, we have
found that for a TCP source which has a medium access bandwidth (between low and high relative
to the burst assembly time), using an adaptive assembly algorithm yields the best throughput
because it can adjust the assembly time to match the TCP congestion control mechanisms.
7.3 LOBS
The generality of the evolving G-MPLS framework makes it a versatile framework for various
underlying switching paradigms. For example, when G-MPLS is applied to OCS in the forms of
MPAS, a wavelength is treated as the label. But such a A- Labelled Switch Path(LSP), which
corresponds to a lightpath, cannot be aggregated at the intermediate node due to the lack of
wavelength merging technology. In order to groom or aggregate traffic carried on different
lightpaths, each lightpath needs to go through O/E/O conversion.
As a natural extension of the G-MPLS in OBS networks, Labelled OBS(LOBS) was proposed
in. LOBS is built upon OBS by letting each control packet carry additional label information. One
of the major benefits of LOBS is to facilitate the seamless integration of IP and WDM by using
IP-based protocols for control while switching data ail-optically. And unlike MPAS, the
association between a label and a wavelength in LOBS is not on the time scale of a connection but

that of each burst, thus making sub-wavelength granularity and statistical multiplexing possible.
Even without wavelength conversion capability, bursts belonging to the same LSP (called LOBS
path) can be sent on different output wavelengths at the ingress node (with a tunable transmitter).
More important, the bursts belonging to different LSPs can interleave on the same wavelength,
that is, bursts arriving on different LSPs (on the same wavelength) can now be merged into an
aggregated LSP.

8 Conclusion
In this report, I have given an introduction to optical burst switching (OBS). Comparison between
this new switching paradigm with other existing optical switching paradigms has been made, and
it has been shown that OBS is not only a cost-effective but also a viable solution for the next
generation optical Internet. We have provided a brief historical review of the early work on burst
switching as well as the state of the art including the prevailing reservation protocol for OBS
networks, called Just-Enough-Time (JET) and described its major features and benefits.
This tutorial has also attempted to provide a comprehensive coverage of research issues
related to OBS. Among the issues covered are various burst assembly algorithms used at the edge
of an OBS network as well as their effect on traffic characteristics of the assembled burst traffic and
in turn the TCP performance. We have also presented various scheduling algorithms as well as
burst contention resolution strategies used in the OBS core. It has been pointed out that
bandwidth-efficient scheduling algorithms like Min-SV can have a fast implementation, and that
burst loss can be reduced using pro-active burst contention resolution algorithms running at the
edge (in addition to the core). Finally, recent work on QoS support, IP/WDM multicast and GMPLS extension have been discussed.
In addition to the challenges in implementing fast and scalable switching fabrics, and related
devices such as FDLs and all-optical wavelength converters, there are many open issues in OBS
architecture research. Chief among them are LOBS path provisioning, protection/restoration
schemes as well as differentiation schemes to combat burst losses due to the inevitable burst
contention in a bufferless OBS network. Others include efficient support of a mixed set of
periodical connections (to emulate SONET) and non-periodical burst transmissions, accurate
single-node as well as end-to-end performance analysis in OBS networks, quantitative cost and
performance comparison between OBS and other switching paradigms, and the design and
evaluation of various TCP implementations over OBS.
OBS has received a lot of attention during the past few years and is fast becoming an
important area of research. This tutorial will hopefully become a useful resource for researchers
working on OBS or those new to this topic.

9 References
[1] D. J. Blumenthal, P. R. Prucnal, and J. R. Sauer, "Photonic packet switches: architectures and
experimental implementations," Proceedings of the IEEE, vol. 82, pp. 1650-1667, November
1994.
[2] G.-K. Chang, G. Ellinas, B. Meagher, W. Xin, S.J. Yoo, M.Z. Iqbal, W. Way, J. Young, H. Dai,
Y.J. Chen, C.D. Lee, X. Yang, A. Chowdhury, and S. Chen, "Low Latency Packet Forwarding
in IP over WDM Networks Using Optical Label Switching Techniques," in IEEE LEGS 1999
Annual Meeting, 1999, pp. 17-18.
[3] M. Yoo and C. Qiao, "Just-enough-time (JET): A high speed protocol for bursty traffic in
optical networks," in Proceeding of IEEE/LEO S Conf. on Technologies For a Global
Information Infrastructure, August 1997, pp.
26-27.
[4] C. Qiao and M. Yoo, "Optical burst switching (OBS)-a new paradigm for an optical Internet,"
Journal of High Speed Networks, vol. 8, no. 1, pp. 69-84, 1999.
[5] J. Turner, "Terabit burst switching," Journal of High Speed Networks, vol. 8, no. 1, pp. 3-16,
1999.
[6] C. Gauger, "Contention resolution in Optical Burst Switching networks," in Advanced
Infrastructures for Photonic Networks: WG 2 Intermediate Report, 2002, pp. 62-82.
[7] Y. Xiong, M. Vandenhoute, and H. Cankaya, "Control architecture in optical burst-switched
WDM networks," IEEE Journal on Selected Areas in Communications, vol. 18, pp. 18381851, October 2000.
[8] F. Masetti et.al., "Design and Implementation of a Multi-Terabit Optical Burst/Packet Router
prototype," in OFC Postdeadline Papers, 2002, pp. FD11-FD13.
[9] http://www.seminarsonly.com